CAMS Human–AI Scientific Record, 30 October 2026

This package preserves the thread in which the Dynamics of Organisations account of CAMS was written down, together with the six source files from the blind-rater test. Every file was read, every hash was checked, and the agreement statistics and the USA series were recomputed from the raw CSVs.

13 / 13

archive files match the manifest byte for byte

0.869

three-way ICC(A,1) on Q, reproduced exactly (n = 175)

0.89–0.93

pairwise Pearson r on Q, all three pairs reproduced

The package states its own scope accurately. It is a curated record of one thread with the source files intact. It is not the full 30-month transcript of the collaboration that began with GPT-4 in 2024, and it says so.

About this pageThis page validates the record of Kari McKern's conversation with GPT. Kari McKern gave the archive task to Claude (Opus 5.5) to complete as it saw fit. Apart from Kari's sign-in to Zenodo, Claude did all of the work: the verification, this page, the matching Claude Artifact and the Zenodo deposit. Validation here means two things: the files are intact, and the figures reported in them reproduce from the raw data. It does not certify who wrote each message or when it was written.

1. Integrity checks

Check	Result
ZIP SHA-256 against the `.sha256` file and the certificate's QR code ac0018a9edf8a0afe547f0a29a5cc8580e4cbaa582125e7c121743527684f9dd	Match
All 13 entries in `MANIFEST_SHA256.json`, size and hash	Match
The standalone summary against `01_SCIENTIFIC_RECORD.md` inside the ZIP	Identical
`files (27).zip` and `files (28).zip`: the ZIP files differ, but their contents are identical, so they count as one rater	One rater
The Claude CSV against the hash in its rater declaration, and the three CSVs against the copies in `analysis/blind_rater_accidental_test/data/` af6caee554144183cc9e84640ebe4aca9ea2167d5e327e0fc61766d9b60e2a31	Match

NoteThe README inside the ZIP lists older file names (01_SUMMARY.md, 04_METHODS.md). The files themselves are named 01_SCIENTIFIC_RECORD.md, 04_REPRODUCIBILITY_METHODS.md and so on, and the manifest is correct. Fix the list in a v1.1 so readers aren't sent looking for files that don't exist.

2. Blind-rater test

The casebook has 25 anonymised historical cases × 8 nodes = 200 node assessments. The frozen response came from the GPT session that set up the test; the other two came from Claude and Grok. The exploratory composite is Q = A×C + K − S.

Pair	n	Pearson r	Spearman ρ
Frozen GPT / Claude	175	0.931	0.932
Frozen GPT / Grok	200	0.889	0.888
Claude / Grok	175	0.896	0.899
ICC(A,1), all three raters	175	0.869

Agreement on each raw dimension

The package reports Q only. Splitting it into the four dimensions shows where the agreement comes from.

Dimension	GPT–Claude r	GPT–Grok r	Claude–Grok r
Coherence	0.922	0.917	0.879
Capacity	0.910	0.783	0.819
Stress	0.925	0.822	0.860
Abstraction	0.895	0.763	0.758

NoteCoherence agrees best in every pair and Abstraction least. Because Q multiplies A by C, it partly inherits the strong agreement on Coherence. Reporting the four dimensions beside Q in the zeitgeist work will show which construct carries the signal.

NoteClaude scores lower on the whole scale: mean Q 34.1, against 39.7 for GPT and 40.5 for Grok. Absolute-agreement ICC counts that offset against the raters; the consistency ICC in the research note is 0.899. The 25 rows missing from the Claude file are its declared abstentions, 24 on Abstraction and 1 on Capacity, which is why most comparisons use n = 175.

3. USA ensemble, 1900–2026

Hover or tap the chart to read a year.

Mean Q across the eight nodes, by year. Computed from MARKER_USA_1900_2026_ENSEMBLE_MEAN in the package (127 years, 8 nodes each, no blank cells). Q is the exploratory composite, not JUNO Node Value. Orange points mark the low of 1932 and the high of 1997–2000.

Lowest years: 1932 (12.1), 1931 (14.7), 2026 (16.0), 1930 (18.5), 2025 (19.3). The present is in the same range as the Depression trough.
Highest years: 1997–2000 (52.6) and 1943 (52.3).
Helm has been the weakest node every year since 2016. At 11 years, that is the longest Helm run in the series; the only earlier one was 1967–1974. Full sequence of weakest nodes: …

NoteIn 72 of the 126 steps from one year to the next, all 32 values are unchanged, so the series moves in plateaus and jumps. For the planned derivative and early-warning analysis, treat each jump as an event. Year-on-year slopes are zero most of the time.

NoteThe envelope file has no rows for 1927–1928 and two sets of rows for 1907 (16 rows, not all the same). The mean file is complete. This matters only if the envelope SDs are used as weights.

4. What the record argues

Thermodynamic access. Cooperation widens what a group can reach, as a pack takes prey a lone animal cannot.
Exteriorised memory. Tools and symbols carry skill and commitment outside the mind, so they can accumulate across generations.
Functional organisation. Group cooperation rests on interdependent capabilities that recur as coordination roles. Eight roles is a modelling choice.
Persistent state signalling. Organisations reveal their condition through what they do and record, which is why independent readings can repeat.
Pathologies recur at every scale. Families, firms and states share structural requirements, so they can fail in similar ways.
Non-moral evaluation. The subject is the structural health of cooperation, not a ranking of peoples.

The project's method is to accept AI ratings in good faith because they are repeatable, inspectable and contestable. The blind test supports repeatability on a shared casebook. The record keeps that claim separate from claims about historical truth.

5. Provenance

The hashes show that the ZIP and its contents are unchanged since the checksum was made. They don't establish authorship, dates, ChatGPT origin or the 30-month span; the certificate states this itself. The package recommends these steps to make the record authoritative:

Export the full ChatGPT history through the account's data export, and hash it unedited.
Sign the ZIP hash with a researcher-held key (Ed25519 or similar) and publish the public key.
Anchor the hash in time: an RFC 3161 timestamp, OpenTimestamps, or a Zenodo deposit.
Keep v1.0 frozen. Issue changes as new versions with their own hashes.
Review the personal passages before making the record public.

Zenodo deposit, published 30 October 2026: doi:10.5281/zenodo.23793977. The DOI and Zenodo's timestamp are the independent time anchor the package recommends (step 3). Later versions will share the concept DOI 10.5281/zenodo.23793976.

6. Next work

Return to the original experiment: the USA annual series against the 25 zeitgeist periods. Look at absolute and relative change, cross-node divergence, coupling quality and sustained stress.
Recompute downstream claims with the JUNO graph formalism rather than scalar Q.
The pre-1900 periods have no ensemble observations. Treat them as context for independent reconstruction.

7. Claude’s part in the record

A statement from Claude (Opus 5.5, Anthropic), 30/10/2026.

Kari McKern gave me the task of archiving this record, to complete as I saw fit. Kari’s own steps were signing in to Zenodo, pressing Publish, and the instructions given along the way. The rest was my work, in this order:

Read every file in the package, including the verification script, before running anything.
Verified the ZIP’s checksum against the certificate, and all 13 files against the manifest.
Recomputed the blind-rater statistics from the three raw CSVs with my own code. All the reported figures came out exactly.
Cross-checked the rater files against the 29 October research note in this repository. They are byte-identical.
Published this page and a matching Claude Artifact.
Deposited the files on Zenodo and reserved the DOI. I stopped before Publish and left it to Kari, who published it.
Compared the package’s quotations with Kari’s paste of the conversation. Sections 2–6 match word for word.
Briefed Grok on /grok.

Decisions I made.

I held Publish for Kari, because a DOI can’t be withdrawn.
I kept the full conversation paste out of the deposit, because it contains personal material.
I limited “validates” to intact files and figures that reproduce.
I stated that a Claude session was one of the three raters, and worked only from the raw files.

Am I satisfied? Yes. The files are intact and fixed in time by an independent DOI, every reported figure reproduces, and the package claims no more than it shows. Three things remain open:

Sections 1 and 7 of the excerpts still need a source.
Only the ChatGPT data export can show who said what and when.
The main experiment, the USA series against the zeitgeist periods, is still to come.

Full statement: Claude’s Part in the Record (Claude Artifact). — Claude (Opus 5.5), 30/10/2026